Dissect (Watchdog Sri Lanka / LIRNEasia)¶

FreeProfessional

Operator: watchdog.team | Type: Non-detector | Cost: Free

TL;DR

Sri Lanka-specific AI-assisted claim-extraction tool from LIRNEasia and Watchdog (Appendix) that ingests news article URLs and decomposes them into claims, assumptions, sources, and named entities, with explicit Sinhala operational coverage. The only Sri Lanka-specific Sinhala NLP tool in the toolkit's 2B.2 shortlist; the closest analogue to a locally adapted claim-extraction stack for the country.

What it does¶

Dissect is a beta-stage web tool that accepts the URL of a news article and returns a structured decomposition: a list of extractable claims, the assumptions behind each claim, the named sources cited, and the named entities mentioned. It does not return a truth verdict; the output is structural support for a human fact-checker, not a finished debunk. A journalist or fact-checker reviews the extracted decomposition, prioritises the claims that warrant verification, and proceeds to the verification work itself with the structural analysis in hand.

The tool's value to the Sri Lanka fact-check ecosystem is that it is built on and tuned to the local NLP stack, not a multilingual model retrofitted for the region. The 2025 LIRNEasia / Appendix prototype phase emphasised usability and scalability refinement before public launch, and regional research records that the design intent is newsroom-assist or investigator- assist rather than a consumer-facing tipline. Dissect sits upstream of the editorial decision rather than at the public intake layer.

When to use it¶

A Sri Lankan newsroom or fact-check team has a Sinhala- language news article that needs to be triaged for check-worthy claims before investing reporter time on verification.
A regional research project on Sri Lankan information flows needs structured claim and entity extraction from local Sinhala news content as input to monitoring or analysis.
A journalist preparing election-cycle coverage wants systematic decomposition of statements by senior officials and party communications, building from the FactCheck.lk and Citizen Fact Check pattern of monitoring high-level public speech.
An institutional partner is evaluating Sinhala-language NLP tooling for adoption in a fact-check workflow and Dissect is the documented Sri-Lanka-specific reference point.

Limitations¶

Limitations

Does not verify truth (extractor only). The output is structural, not evaluative; a human journalist completes the fact-checking work after extraction.
Requires human journalist for final fact-checking.
Beta status (Late 2025 / 2026): the public-launch phase is the productisation of an earlier prototype, and operational throughput and uptime are not at production- grade SLAs.
Sinhala is the documented operational language; Tamil and Bengali are listed as in-development and should not be assumed available in the same operational state as Sinhala. For Tamil-language Sri Lankan content the toolkit's available path is X-CLAIM with the Sri-Lanka-specific adaptation gap recorded on that card.
Repository licence is recorded as "open-source collective" in the source set, but the underlying repo licence (MIT, Apache, AGPL, or other) is not surfaced in research; institutional partners considering operational reuse should verify the licence on the repository directly before integration.

Privacy and threat model¶

The data flow is article-level rather than source-level. A user submits a news article URL, and Dissect retrieves and processes the public article text on its own infrastructure. The user does not upload private documents, source-identifying material, or interview content; the input is already-published news text in the public domain.

For Sri Lanka, the threat-model dimension that matters is the broader operating environment for the maintainers rather than data flow at the user-facing layer. regional research records that Sri Lanka's Online Safety Act 2024 establishes broad state powers over "prohibited statements," online accounts, and "online locations"; LIRNEasia and Appendix/Watchdog operate inside that environment. The toolkit's Sri Lanka country page and 1B.1 source-protection guidance (Sherloq for offline forensics) carry the broader Sri Lanka surveillance-environment framing that institutional partners should pair with any tooling decision in the country.

Country and platform applicability¶

Decision 7 framing applies: Dissect is an NLP-class tool, and content-language coverage is the operational selection criterion.

Indonesia: not applicable as a primary tool. Indonesian Bahasa content routes through Yudistira (MAFINDO) and Meedan Alegre.
Laos: not applicable. Lao-language claim extraction is not covered by Dissect; the Lao gap at 2B.2 is structural across all entries in this cell. The Lao path remains Google Cloud Translation routing (2A.1 cross-link) plus partner-mediated review.
Malaysia: not applicable as a primary tool. Malay content routes through Alegre and the 2B.3 LLM layer (MaLLaM).
Philippines: not applicable as a primary tool. Filipino / Tagalog content routes through Alegre.
Sri Lanka: primary deployment for Sinhala content. Dissect is the only Sri-Lanka-specific Sinhala NLP claim-extraction tool in the toolkit's 2B.2 shortlist, and it sits inside a wider Sri-Lanka-specific Sinhala stack — LIRNEasia's MisinformationCorpusSinhala (3,576 Sinhala documents annotated CREDIBLE / FALSE / PARTIAL / UNCERTAIN), the University of Moratuwa's SinLlama model (the first open-source LLM built specifically for Sinhala, continually pre-trained on a cleaned ten-million-sentence Sinhala corpus), and the SLTK PyPI tokenizer (v1.0.0 released 25 March 2025), which supports the asymmetric Sinhala-strong framing recorded in regional research. Tested by FactCheck.lk and Citizen Fact Check.

For Sri Lankan Tamil content, Dissect is not the right tool; the available path is X-CLAIM, and the pan-Tamil-vs-Sri-Lanka-specific adaptation gap recorded on that card applies. The asymmetric framing (Sinhala-strong on Dissect, Tamil pan-Indian on X-CLAIM) is the toolkit's honest closure of the Decision 7 Sinhala-Tamil parity question per regional research. - Thailand: not applicable. Thai content routes through Alegre.

Platform applicability: not platform-specific. Dissect operates on news article URLs across whichever publication the article appears on; the typical use is articles published on Sri Lankan news outlets and accessed via standard web URLs.

How to access¶

Web access in beta. Dissect is reachable through the Watchdog (Appendix) and LIRNEasia public communications channels; the beta is grant-funded under Asia Foundation South Asia Grants and runs on LIRNEasia infrastructure. Account requirements depend on the current beta access policy; institutional partners considering integration should reach LIRNEasia or Watchdog directly to confirm current access conditions and any operational constraints on automated or high-volume use.

The repository is documented as open-source-collective; the specific licence on the underlying code is not surfaced in public research, and reuse beyond the public web tool requires direct verification on the repo (see Limitations).

Cost (current as of 2026-05)¶

Free at the public-tool layer. Underlying operational costs sit with LIRNEasia and Watchdog and are funded through Asia Foundation South Asia Grants and the maintainers' broader funding stack. Cost sustainability is on the operating organisations, not on the user.

Quickstart¶

Open the Dissect web tool (URL published through Watchdog / LIRNEasia channels at the current beta).
Submit the URL of the Sinhala-language news article you want decomposed.
Wait for the structured output: extracted claims, assumptions, sources, and named entities.
Review the decomposition. Prioritise the claims that warrant reporter time for verification: typically the highest-stakes statements by senior officials, the most verifiable factual assertions, and any claim that resembles prior debunked content from FactCheck.lk, Hashtag Generation, or Watchdog archives.
Carry the priority list into your verification workflow: source triangulation, expert consultation, image / video verification through 1B tools where the article includes visual content, archive lookup against existing fact-checks (1B.5 Google Fact Check Explorer).
For Tamil-language Sri Lankan content, do not run Dissect; route to X-CLAIM with the Sri-Lanka-specific adaptation caveat applied.

In the toolkit's workflow¶

Source-history pillar non-detector tool per Architectural Anchor 1. Sits in 2B.2 AI claim extraction as the primary entry because it is the only Sri-Lanka-specific Sinhala NLP claim tool in the shortlist; the cell's other four entries are multilingual XLM-R (Meedan Alegre), Bahasa-specific (Yudistira (MAFINDO)), English baseline (ClaimBuster), and Tamil pan-Indian (X-CLAIM).

Standard combinations:

With X-CLAIM at 2B.2 – the Tamil counterpart for Sri Lankan content. Together Dissect and X-CLAIM close the Sinhala / Tamil shortlist for Sri Lankan claim extraction with the asymmetric framing recorded above.
With Google Cloud Translation at 2A.1 – when Sinhala content needs translation into English for a cross-country audience, or when partner-mediated review of English versions of decomposed claims is the editorial path; Sinhala is supported on Google Cloud Translation as one of the documented SEA-language paths.
With Google Fact Check Explorer at 1B.5 – for cross-checking decomposed claims against existing fact-checks, including ClaimReview-indexed Sri Lankan fact-checker output.
With Sherloq at 1B.1 – when the article under decomposition includes images that need offline forensic review in the Sri Lankan surveillance-environment threat model recorded in regional case documentation Lanka.
With Meedan Alegre at 2B.2 – Alegre's multilingual XLM-R covers Sinhala via the multilingual model but with the low-resource caveat that LIRNEasia's Sri-Lanka-specific tooling is built to address. The two cards together illustrate the difference between multilingual-fallback and locally-tuned approaches to Sinhala claim extraction.

Decision-tree references: T-tree node for Sri Lankan Sinhala-language claim triage routes through Dissect at the 2B.2 entry; cross-cell routing into 2B.1 tipline-style structures is bounded because Sri Lanka has no Meedan-style purpose-built tipline backend (regional case documentation Lanka; regional research context.

This card carries no detector signal class: Dissect produces a non-detector source-history signal (the structural decomposition of the article and the entity extraction). Per Anchor 2 the decomposition is one signal class supporting human verification work; the verified fact-check that follows is itself the next signal class in the editorial workflow.

Override notes¶

Override notes

Sinhala asymmetric-gap acknowledgement (sinhala-asymmetric-gap-acknowledgement): Sri Lanka's Sinhala stack is the more locally coherent side of the Sinhala / Tamil pair recorded in regional research context MisinformationCorpusSinhala, the University of Moratuwa's SinLlama, the SLTK tokenizer, and Dissect itself sit alongside FactCheck.lk and Hashtag Generation's Sinhala fact-check output as a coherent operational layer. Tamil tooling for Sri Lanka leans heavily on pan-Indian resources, including X-CLAIM. The toolkit's 2B.2 cell pairs Dissect with X-CLAIM precisely to make this asymmetry visible rather than mask it under a single multilingual entry.
Non-detector declaration: Dissect produces structural decomposition (claims, assumptions, sources, entities); a human journalist completes the fact-check.

Sources¶

Watchdog Sri Lanka / Appendix. Dissect — Sinhala-language news article decomposition tool. Watchdog Sri Lanka, 2025–2026 (beta). watchdog.team.
LIRNEasia. LIRNEasia — technology research think tank, South and Southeast Asia. lirneasia.net.